Using Hybrid Physics-Informed Neural Networks for Digital Twins in Prognosis and Health Management
Motivation
- Simulations are pervasive in every domain of science and engineering, but they often have constraints such as large computational times, limited compute resources, tedious manual setup efforts, and the need for technical expertise. Neural networks not only accelerate simulations done by traditional solvers, but also simplify simulation setup and solve problems not addressable by traditional solvers.
- NVIDIA Modulus is a physics-informed neural network (PINN) toolkit for engineers, scientists, students, and researchers who are getting started with AI-driven physics simulations. You may also be looking to leverage a powerful, existing framework to implement your domain knowledge and solve complex nonlinear physics problems with real-world applications.
- A success story of a Modulus application today is in the use of hybrid PINNs for digital twins in prognosis and health management.
Case
Aircraft use case study
Motivation
- Maintenance of engineering assets and industrial equipment (such as aircraft, jet engines, wind turbines, and so on) is critical for safety as well as enhanced profitability in the services and warranties of these assets. Effective preventive maintenance requires the knowledge of the various operating parameters and their impact on the wear and tear of equipment. Simulations, advanced analytics, and deep learning algorithms enable the predictive modeling of complex systems and their operating environment.
- In this research project, Prof. Viana and his team of researchers built predictive models for fatigue crack growth prognosis on aircraft window panels (Figure 1), where models were trained using historical flight records (origin and destination airports, cruise altitude, and so on) and limited inspection observations (such as the crack length data for only a portion of the fleet, and so on
challenges
- There are two main challenges.
- First, the data is highly unbalanced. For the aircraft fleet that was being analyzed, there were only 182,500 input points and only 25 output points. Building purely machine learning models under such circumstance is extremely hard.
- Second, while the conventional physics-based models maybe accurate, they often entail engineering assumptions regarding loading conditions. Considering that the sample size for this problem includes a fleet of 500 aircraft, these simulations must be done a few million times.
Main work
- Full-fledged finite element analysis configured for crack growth simulation is extremely expensive. Therefore, it is simply not feasible for digital twin applications. This is true even if we were to run simulations for hundreds of aircraft many times over, as we optimized inspections and decided how to swap routes for a few aircraft. A parameterized physics-driven AI model is constructed in Modulus that satisfies the governing laws of linear elasticity, as follows:
σ j i , j + f i = 0 , σ i j = λ ϵ k k δ i j + 2 μ ϵ i j , ϵ i j = 1 2 ( u i , j + u j , i ) \sigma_{j i, j}+f_{i}=0 ,\sigma_{i j}=\lambda \epsilon_{k k} \delta_{i j}+2 \mu \epsilon_{i j}, \epsilon_{i j}=\frac{1}{2}\left(u_{i, j}+u_{j, i}\right) σji,j+fi=0,σij=λϵkkδij+2μϵij,ϵij=21(ui,j+uj,i)
其中, σ i j \sigma_{i j} σijis the Cauchy stress tensor, δ i j \delta_{i j} δijis the Kronecker delta function and ϵ i j \epsilon_{i j} ϵij is the strain tensor.
Method
- The inputs to this parameterized model are the loading condition (the hoop stresses) and the spatial coordinates of batches of a point cloud within the computational domain. The outputs are the stresses and displacements. The network architecture consists of a Fourier feature encoding layer followed by several fully connected layers.
- Several techniques are used to enhance the accuracy and convergence speed of the model, such as network weight normalization, signed distance loss weighting, differential equation normalization and nondimensionalization, and XLA kernel fusion
Result
来源:
https://developer.nvidia.com/blog/using-hybrid-physics-informed-neural-networks-for-digital-twins-in-prognosis-and-health-management/